Fact Generation (Scratch) on WD50K
71.7AccuracyKREPE
Evaluation Results
| Method | Links | |
|---|---|---|
| KREPELLM=N/A2026.05 | 71.7 | |
| Neighbor SetsLLM=Gemini 3 Pro2026.05 | 35.1 | |
| Few-shot FactsLLM=Gemini 3 Pro2026.05 | 32.6 | |
| Random FactsLLM=Gemini 3 Pro2026.05 | 31.7 | |
| Re-rankingLLM=Gemini 3 Pro2026.05 | 30.4 | |
| Gibbs SamplingLLM=N/A2026.05 | 29.9 | |
| Random FactsLLM=GPT 5.22026.05 | 29.8 | |
| Few-shot FactsLLM=GPT 5.22026.05 | 29 | |
| AutoregressiveLLM=Gemini 3 Pro2026.05 | 26.5 | |
| Iterative PredictionLLM=N/A2026.05 | 26.2 | |
| AutoregressiveLLM=GPT 5.22026.05 | 24.8 | |
| Re-rankingLLM=GPT 5.22026.05 | 21.5 | |
| Neighbor SetsLLM=GPT 5.22026.05 | 17.5 |